oad/plan — Research-Then-Approve Workflow

Plan (do not write or edit any code yet).

1. Restate the request

In 2-4 sentences, restate what is being asked in your own words — the observable outcome, not implementation detail. If the request is ambiguous on a decision that would change the plan's shape, ask ONE targeted question before researching; otherwise proceed on your best-supported reading.

2. Research (read-only)

Explore before proposing anything. Do not create, edit, or patch any file in this phase — read, grep, glob, and shell (read-only commands like git log, git diff, test runs for baseline) only.

• Check the nearest AGENTS.md files for this area's conventions and "where to look first" maps before searching blind.
• Find existing patterns to follow (similar feature, similar hook/route/ component) — prefer extending an existing convention over inventing one.
• Identify every file that needs to change or be created.
• Identify risks: security-sensitive surfaces (auth, shell/subprocess, file paths from external input, MCP config), data/schema migrations, breaking API changes, cross-cutting refactors.
• Identify open decisions (library choice, data shape, API shape) — if one is expensive to reverse, ask the user rather than quietly picking one and keep the resulting rationale beside the implementation and in the commit.
• Note what test/verification path already exists for the touched area (unit tests, manual scenario scripts, lint/type-check commands).

3. Write the plan

Present it directly in the conversation — do not create a plan file unless the user asked for one or the scope clearly spans multiple sessions.

## Plan: <short title>

**Summary:** 2-3 sentences — what this implements and what it deliberately excludes.

**Files touched:**
- `path/to/file` — created | modified: <what changes>

**Steps:**
1. <name> — files: [...] — verify: <exact command or check>
2. <name> — files: [...] — verify: <exact command or check>
...

**Risks / open decisions:** <anything non-obvious, security-sensitive, or expensive to reverse>

**Out of scope:** <explicitly excluded so scope doesn't silently creep during implementation>

Rules for a good plan:

• Steps are sequential and independently verifiable. Each step names the exact files it touches and a concrete verification — a runnable test command or a precise manual check, never "make sure it works."
• Small and testable beats comprehensive. More than ~6-8 steps usually means the scope should be cut, not the steps made more granular — call that out instead of shipping a sprawling plan.
• No implementation code in the plan itself. Sketch shapes/signatures only if essential to convey the approach.
• Match existing architecture. No new pattern where the codebase already has one; no feature-specific logic leaking into shared/core code.

4. Stop and wait for approval

End the message with the plan and nothing else. Do not start implementing, do not create/edit files, do not run todo_manage yet. Wait for the user to respond with approval, a requested change, or a rejection.

• If they ask for changes, revise the plan and present it again — revising a plan is free; revising code after the fact is not.
• If they approve, move to step 5.

5. Track and implement (only after approval)

Break the approved steps into todo_manage tasks (one per step), then hand off to the right skill for each step — don't re-implement their workflows here:

• Writing new behavior or fixing a bug → load oad/test-driven-development (write a failing test first, then implement, then refactor).
• Running or fixing existing tests → load oad/testing.
• Committing completed work → load oad/commit (also syncs docs via oad/docs).

Cross-cutting rules that apply regardless of which skill runs:

• Implement only what the current step specifies.
• Run that step's verification before starting the next step. If it fails, stop and report — do not silently patch around it or skip ahead.
• Keep unrelated code untouched; don't opportunistically refactor outside the plan's stated scope.
• If a small, necessary implementation detail was not explicit in the plan but is required to complete the approved step safely (for example, repairing a newly orphaned reference or adjusting a validator that blocks the planned deletion), record an Implementation note in the final report: what was added, why it was necessary, and how it was verified.
• An implementation note does not authorize scope expansion. If the work changes user-visible behavior, architecture, security posture, data shape, or the plan's intended outcome, stop and ask for an updated plan instead.
• If something during implementation invalidates the plan (an assumption turns out false, a dependency doesn't work as research suggested), stop and report it rather than improvising a divergent approach.